Kihagyás

Crystallization protocol#

Origin: Originally written in Hungarian as part of MyForge Vault 11.11 — Superintelligent Vault project. Source: Crystallization-protocol (Hungarian version).

When a session closes (/11.11stop), the agent must transform learnings into long-term knowledge: every ## Learnings bullet is propagated into the appropriate persistent layer. This is not optional. The vault only "compounds" (in the Karpathy sense) if knowledge is carried into the persistent layers at every session close.

Protocol overview (manual mode — original)#

session close → agent collects learnings → classifies them
              → BATCH PREVIEW: shows them all to the user at once
              → user OKs / modifies / skips
              → agent executes
              → propagation log appended to the session file

Auto-mode addition (G-Eval scoring layer)#

In addition to the manual protocol, an automatic scoring layer can be added that computes a 4-dimensional G-Eval confidence for every Learning bullet. The routing decision tree is unchanged — scoring appears as a signal, NOT as a replacement for human decision at conservative (0.95) and aggressive (0.85) thresholds.

session close → crystallize <slug> --scorer claude-code --with-context  (opt-in)
              → Phase 1: pending request file
              → general-purpose Agent scores using G-Eval prompt (4 dim × 5 scale, with corpus context)
              → Phase 2: response file → confidence router → audit log
              → routing decision tree (unchanged)
              → batch preview to user (now with confidence color coding)
              → user OK → propagation → propagation log

Confidence routing:

Confidence Route Threshold rules
≥ threshold auto-prop Shadow (1.0): audit-log only · Conservative (0.95): 1-shot user confirm · Aggressive (0.85): no confirm
0.70 — threshold batch-preview To user per decision-tree routing
< 0.70 discard Audit log only, NOT propagated

Threshold config: held in a hot-reloadable file. Default 1.0 (shadow); use 0.95 (conservative) once a calibration benchmark shows verdict agreement > target (e.g., 96.7% vs 90% goal).

Cost: $0 (claude-code scorer via subagent-fanout, see claude-code-subagent-fanout).

Routing decision tree#

For each Learning bullet, the agent walks through this list; first match wins:

  1. Architecture-level decision ("we choose X over Y because…", with reasoning + alternatives) → new ADR document
  2. Vault convention / rule (new tag, new frontmatter field, new folder rule) → tag taxonomy or frontmatter schema
  3. New evergreen concept / playbook (a pattern, a recipe, a process) → new wiki entry
  4. New abbreviation / slug / codename → glossary
  5. Server / port / cron / DB / service knowledge → infrastructure memory
  6. Skill description (agent skill / capability) → skill map
  7. User preference (style rule, "always use X in section Y") → user preferences memory
  8. Dashboard / access rule → access control memory
  9. Project-specific knowledge (status change, gotcha, port value, deploy recipe for one project) → project file
  10. New TODO / task → backlog (correct priority section)
  11. Else → ask the user

Batch preview UX#

The user wants to see all suggestions at once, NOT bullet-by-bullet:

🧠 N learnings to propagate — I suggest:

[1] "<bullet 1 quote>"
    → infrastructure-memory ▸ "Next.js gotchas" section (new subsection)
    → preview: <first 2 lines of the text to be inserted>

[2] "<bullet 2 quote>"
    → user-memory ▸ "UI/UX preferences"
    → preview: ...

[3] "<bullet 3 quote>"
    → new ADR: <YYYY-MM-DD topic>.md
    → status: accepted, affects: <project>
    → preview: ...

[4] "<bullet 4 quote>"
    → new evergreen: wiki/<topic>.md
    → preview: ...

OK? You can reply briefly:
- "OK" → all executed
- "1-3 OK, 4 to project-X.md" → 1-3 fine, route 4 elsewhere
- "skip 2" → drop 2, keep the rest
- "stop" → nothing

Propagation log#

A ## Propagation log section is appended to the session file, time-stamped:

## Propagation log

- 2026-04-30T12:34 — [1] → infrastructure-memory (new gotcha)
- 2026-04-30T12:34 — [2] → user-memory (UI/UX prefs added)
- 2026-04-30T12:34 — [3] → new ADR <YYYY-MM-DD topic>.md
- 2026-04-30T12:34 — [4] → new wiki/<topic>.md
- 2026-04-30T12:35 — Backlog updated (3 new TODOs #project/<slug>)

This is auditable later: you can trace "where did this learning come from" → the propagation log points back to the source session.

What NOT to propagate#

  • Transient state that will be irrelevant in 1 day ("the build was running on the server just now")
  • Bug info we already fixed (the fix is in the code, no document pollution needed)
  • Implementation steps that were only interesting during the session ("I tried X, it didn't work")
  • Duplicate knowledge already documented elsewhere (then just link)

If unsure → keep it in ## Learnings (lives in the session file as raw) but DO NOT propagate to persistent layers.

Edge cases#

Situation What to do
A bullet fits in two places Pick one, or duplicate with a short link (put the source in the ADR/wiki, only → see [[…]] elsewhere)
No bullet is propagation-worthy OK, just write "## Propagation log\n\n- (no propagation-worthy learnings)"
User says "stop" Respect it, just record "user skipped" in the propagation log
User changes the target Override the suggestion, propagate per the user's redirection
New ADR/wiki needed, but user is offline Leave in the session ## Learnings with a flag: <!-- TODO-PROPAGATE: new ADR or infra memory --> — handle next time

Combined with aggressive context loading#

If you have already preloaded the project file + last 5 sessions + ADRs at session start (per the context-loading protocol), then propagation is faster: the agent knows which section the new knowledge fits into and won't duplicate.

Audio overview#

  • EN narration (Charon voice): [[.vault-nb/audio-overviews/Crystallization-protocol.en.mp3]]
  • HU narration (Kore voice): [[.vault-nb/audio-overviews/Crystallization-protocol.hu.mp3]]

Generated via Gemini 3.1 Flash TTS preview. ~1-2 minutes each. See gemini-3-1-flash-tts-pipeline for the pipeline.